You are on page 1of 27

Math. Proc. Camb. Phil. Soc.

(1975), 77, 43 MPCPS 77-5 Printed in Oreat Britain

43

Spectral asymmetry and Riemannian Geometry. I
B Y M. F. ATIYAH, V. K. PATODI AND I. M. SINGER Oxford University Tata Institute for Fundamental Research, Bombay Massachusetts Institute of Technology, Cambridge, Massachusetts (Received 25 March 1974) 1. Introduction. The main purpose of this paper is to present a generalization of Hirzebruch's signature theorem for the case of manifolds with boundary. Our result is in the framework of Riemannian geometry and can be viewed as analogous to the Gauss-Bonnet theorem for manifolds with boundary, although there is a very significant difference between the two cases which is, in a sense, the central topic of the paper. To explain this difference let us begin by recalling that the classical GaussBonnet theorem for a surface X with boundary Y asserts that the Euler characteristic E(X) is given by a formula:

where K is the Gauss curvature of X and cr is the geodesic curvature of Y in X. In particular if, near the boundary, X is isometric to the product Y x R+, the boundary integral in (1-1) vanishes and the formula is the same as for closed surfaces. Similar remarks hold in higher dimensions. Now if X is a closed oriented Riemannian manifold of dimension 4, there is another formula relating cohomological invariants with curvature in addition to the Gauss-Bonnet formula. This expresses the signaturef of the quadratic form on H2(X, R) by an integral formula (1-2) where px is the differential 4-form representing the first Pontrjagin class and is given in terms of the curvature matrix R by px = (2n)-2TrR2. I t is natural to ask if (1-2) continues to hold for manifolds with boundary, provided the metric is a product near the boundary. Simple examples show that this is false, so that in general O. (1-3)

The notation in (1-3) is meant to emphasize t h a t / ( F ) depends only on Y (as an oriented Riemannian manifold) and not on X. In other words, if X' is another manifold with boundary Y, then s i g n X - - | 2>1 = s i g n X ' - - |
V

p'v
P+Q

•f The signature of the quadratic formes? — S 3^ ia p — q.
1 p+1

44

M. F. ATIYAH, V. K. PATODI AND I. M. SINGER

This is an easy consequence of (1 -2) applied to the closed manifold l u ( - Z ' ) obtained by glueing X' (with orientation reversed) to X along Y. We are therefore left with the question of understanding how f(Y) depends on the metric on Y. The following properties are clear from (1*3): / is a continuous function of the metric, f(-Y) = -f(Y), (1-4) (1-5)

where — Y is Y with the same metric but with opposite orientation. A natural conjecture, consistent with (i) and (ii) would be that/is given by some integral expression f(Y)= I 6, where 6 is a 3-form on Y canonically constructed out of the metric. J F However, if this were the case, it would imply that / was multiplicative for finite coverings, namely f(Y) = df(Y), whenever F is a d-fold covering of Y. Explicit examples show that this is false: in fact, for Y = 3-sphere,/(Y) = 0 but/(F) + 0 for a suitable lens space Y (a quotient of the 3-sphere by the cyclic group of order d)f. Thus / must be a global invariant of the metric. Rather surprisingly, and this is our main result, it turns out to be a spectral invariant. We recall that the eigenvalues A of the Laplace operator of the metric are conveniently studied via the Zeta-function:{ £(«) = S A-8. (1-6) Real-valued invariants of the metric satisfying (1-4) can then be obtained by evaluating £(s) at some point where it is known to be finite. Invariants obtained in this way are however insensitive to the choice of orientation so that they will not satisfy (l-5). To find a function of the eigenvalues which satisfies (1-5) we have to do something more subtle. To begin with, we should use the Laplace operator on forms as well as on scalar functions. Next we observe that this total Laplace operator A is in fact the square of a self-adjoint first order operator§ B = ± (d* — *d). Thus the eigenvalues of A are of the form A2, where A is an eigenvalue of B. Now, unlike A, the operator B is not positive so its eigenvalues can be positive or negative. Taking this sign into account, we can therefore refine (1-6) to define a new function
V(s)=

2(signA)|A|-.
A#0

(1-7)

If we regard (1-6) as a generalization of the Riemann Zeta-function, then (1-7) can be regarded as analogous to the Dirichlet Z-functions. Since B involves the * operator (linearly), it follows that reversing the orientation of Y changes B into —B and hence 7](s) into —rj(s). If we now evaluate TJ(S) at a suitable value of s, we will therefore obtain an invariant satisfying (1-5). To decide what value of s is a reasonable candidate, we consider the behaviour of our function / with respect to scale changes of the metric. Since a scale change — • &2gri}-, k a positive constant) does not affect the curvature, it does not alter the (git i
f These examples will be treated in Part II. % An eigenvalue is repeated in (1-6) according to its multiplicity. § Signs are technically most important in what follows, but to avoid complications at this stage we simply write + . The correct signs will be found in section 4.

in the classical situation discussed in the previous paragraph. Of course it is implicit in what we have said that TJ(S) has a finite value at s = 0. the 5/-functions must behave in a different way. However. for the operator B above. To sum up. Note that in the classical notation for flat space BeT is essentially the operator /0 div \ \grad curl/ acting on the 4-vector consisting of a function and a vector-field (which may be identified with a 2-form). This turns out to have a non-local character which is related to the global topological invariant studied in Reidemeister torsion. r/(s) has at first sight a simple pole at s = 0 with an explicitly computable residue. Euler characteristic and arithmetic genus. because B is clearly the direct sum of two operators Bev. On the other hand an easy check (see ((2). just as for £(s). the integral formula for this residue vanishes identically so that 97(0) is finite. that in computing TJ(S) the only relevant part of the operator B is d* acting on dQ2*"1. we show. which are given by explicit integral formulae. Now. the differences between the three cases are substantial and we do not see at present how to unify them under any common generalization. The only value of s for which 7j{s) is unaltered is therefore s = 0.Spectral asymmetry and Riemannian geometry. In fact everything goes through for 4&-dimensional manifolds (as stated in Theorem (4-14)). Thus we can remove the factor \ if we replace the operator B by the operator 5 e v . The behaviour near s = 0 of the two functions is however somewhat different. section 5)) shows that the eigenvalues A of B become &~XA and so t](s) becomes k~srj(s). In view of Theorem (4-14) and of the non-local character of/. Each of these three papers introduces an analytic invariant for manifolds which is related to a classical index invariant: namely the signature. The explanation for this difference between r/ and f is that r/ involves the separation of Spec B into positive and negative. only curl contributes to the computation of ij(s). in Proposition (4-20). I 45 value of/. In fact our main result (Theorem (4-14)) asserts t h a t / ( F ) = ^rj(O). this value is defined by analytic continuation in s from the half-plane of absolute convergence (Re s large). Whereas £(0) is finite and explicitly computable as an integral. In the above discussion. Moreover. . I t is a remarkable fact that. This situation has some close analogies with the analytic i?-torsion studied in (14) and (15). the main difference being that the simple integrand \px in (1 • 3) has to be replaced by the Hirzebruch i-polynomial in the Pontrjagin forms in the general case. For positive self-adjoint elliptic operators the Zeta-functions given by (1-6) have finite values at s = 0. From yet another point of view. which is roughly in the middle dimension. Thus. this paper has something in common with (14) and (15). Bodd acting on forms of even and odd degrees respectively and * switches the two so that the ^/-invariant of Bey is half the ^-invariant of B. This is a global operation and it accounts for the non-local character of y. The factor \ has an obvious explanation. There the appropriate ^-function turns out to vanish at s = 0 and one then proceeds to consider the next term in the Taylor expansion. we restricted ourselves to 4-dimensional manifolds X (with 3-dimensional boundary Y) only for simplicity. we see that the simplest spectral invariant of Y that behaves like/( Y) is 1/(0). namely £'(0).

afiniteindex. It is natural therefore. SINGER Thefinitenessof 7/(0) turns out to be a general phenomenon valid for all self-adjoint elliptic operators. This is the topological obstruction referred to above which shows that there is no (elliptic) boundary condition of local type for the signature operator. For our global boundary condition P{(j>\ Y) = 0 however. This index can be identified with the signature of X so that Theorem (4-14) appears as an index formula. in particular. for a manifold X with boundary Y. then it turns out that we can indeed set up an appropriate boundary value problem. PATODI AND I. . o~ is a bundle isomorphism and B is the self-adjoint operator on Y described earlier. K.e. If P denotes the orthogonal projection onto the space spanned by the eigenfunctions with A ^ 0. thisfinitenesscan be proved in exactly the same manner as for the particular operator B. i. In fact there is an operator (the ' signature operator') acting on certain spaces of differential forms whose index can be identified via Hodge theory with the signature of the manifold. or alternatively one deduces it fom the analogue of Theorem (4-14). That is. It turns out that^> is not deform able (through idempotents of rank \m) to a matrix function of y alone. If we enlarge our point of view however and permit global boundary conditions. to try to set up a suitable elliptic boundary value problem for the signature operator whose index will be the signature of X. we have a good elliptic theory and. For general self-adjoint operators. F. we require that the boundary value </> \ Y should lie in the subspace spanned by the eigenfunctions (j>k of B with A < 0. such a direct argument seems difficult and instead we shall deduce it (in Part III) from the special Riemannian cases by topological methods. to the symbol of a multiplication operator. M. In section 3. ATIYAH. Near the boundary. We now return to explain the method of proof of our main theorem. £) is an idempotent mxm matrix of rank \m defined on the cotangent sphere bundle of Y. For our boundary condition. the boundary condition is P{(j>\Y) = 0. we derive a general •(• At least for odd-dimensional manifolds.f For certain other operators arising in Riemannian geometry (operators of 'Dirac type'). one an integral over the interior X and the other 1/(0) coming from the boundary Y. V. the index being expressed as the sum of two terms.46 M. one computes the residue as an explicit integral and then shows it vanishes identically. the signature operator takes the form (1-8) where u is the normal coordinate. The difficulty with this programme is that there are topological obstructions to the existence of such boundary conditions (1) and these obstructions are non-zero for the signature operator. This is on the assumption that we are looking for classical or local boundary conditions as for example in the Dirichlet or Neumann problems. The operator P is pseudo-differential and its symbol p = p(y. It is clear that the right context in which to view Theorem (4-14) is therefore that of index problems for such global boundary conditions. Recall first that the Hirzebruch signature theorem for closed manifolds is a special case of the general index theorem for elliptic operators.

there are some 3-manifolds arising in the theory of real quadratic fields for which our invariant 7/(0) turns out to coincide essentially with the value L(0) of a certain . we consider in some detail the case of Riemann surfaces with boundary. In the second half of section 4 we show how the special case of Theorem (4-14) fits into the general Riemannian framework. denned by (1-3). Y) in the absolute cohomology H*(X). another important clue was provided by a joint (unpublished) study of the first author and G. and so is certainly squareintegrable. If however our interior operator is one of the classical' Dirac type' operators arising in Riemannian geometry then the main results of (2) enable us to identify the integral with an explicit expression in the curvature. in which the absolute and relative cohomology groups of X and (X. The result we need is. Instead we have to connect cohomology to the null space JV of the signature operator with its global boundary condition. I 47 analytical result. In denning our ^-function by (1-7). In particular. The main point here is to carry out the analogue of the Hodge theory for manifolds with boundary. then. This image is so to speak the part of the cohomology that lies half-way between the absolute and the relative. To explain our result. Unfortunately this is not what we need here. This study was concerned with a periodic family of self-adjoint elliptic operators and . One of the primary motivations of our present work was in fact an attempt to understand the significance of Hirzebruch's result in the wider context of Riemannian geometry. it is naturally acted on by the duality operator *. Note that a form in Jf extends to an exponentially decaying harmonic form on X. and we shall treat these in detail in Part II of this paper. and in fact the attempt to use these classical boundary conditions held up progress for a long time. the interior integral is only given implicitly since it depends on the asymptotic behaviour of the corresponding heat kernel. Lusztig. some of which have been mentioned in ((5). Now there is already such a treatment (8).L-function of thefield. used successfully in (14). and when represented by harmonic L2-forms on X. This is done in section 4 and leads to results such as Theorem (4-2) for the Dirac operator on spinors. particularly for totally real number fields. Conversely we hope that our results may have Number theoretical significance. In the search to identify the invariant/(F). giving such an index formula for first-order systems which are of the simple type (1-8) near the boundary (with B now being any selfadjoint first-order elliptic differential operator on Y). Y) are computed in terms of harmonic forms satisfying appropriate local boundary conditions. it is convenient to introduce the non-compact' elongation' X of X obtained by attaching the semi-infinite cylinder Y x R~ to the boundary Y of X. In Theorem (3-10).Spectral asymmetry and Riemannian geometry. Theorem (3-10). we shall study the relation with coverings. We shall also discuss the relation of our invariant with the recent work of Chern and Simons (6). In fact. exploiting the non-multiplicative character of TJ(O) alluded to earlier. We shall return to this question in a future publication. As a simple illustration. section 2). we mentioned the analogy with the L-functions of Number Theory. There are many interesting generalizations and applications of our results.Modulo this identification Theorem (4-14) for these 3-manifolds reduces to a result of Hirzebruch(lO). Proposition (4-9) asserting that the space of square integrable harmonic forms on X is naturally isomorphic to the image of the relative cohomology H*(X.

The fundamental solution of this cylinder problem then provides us in section 3 with a parametrix near the boundary for our boundary value problem on the compact manifold X. Combined in the usual way with an interior parametrix. In this section. Then P is a pseudo-differential operator. Jr . E) -> C°°( 7. K. E a vector bundle over Y and A: Cx( Y. u) of E lifted to Y x R + (which we still denote by E). The precise relation will be explained in Part III. we shall make some explicit calculations which will be basic to the rest of the paper. Let Y be a closed manifold. ending up with formula (2-25). 2. Clearly D is elliptic and its formal adjoint is D*=-l + A. V. Computations on the cylinder. Roughly speaking the spectral flow is the net number of eigenvalues that change sign (from — to +) while the parameter of the family is completing a period. This is quite adequate for finiteness and regularity theorems but it does not deal with the index which must refer to a determined system. This can be treated quite explicitly assuming only well-known results concerning elliptic operators on the closed manifold Y.% Then A has a discrete spectrum with real eigenvalues A and eigenfunctions <f>x. but we have preferred to give a direct elementary treatment. this special flow clearly has something in common with our 7/-invariant. Since it involves contrasting the positive and negative eigenvalues. so that the inner product for sections of E is given as usual by U(y)> 9(y)) dy. Clearly P = %B~1(B+ \B\). The only difference now is that our heat operator must take into account the boundary condition and this produces a boundary contribution in the asymptotics. SINGER an integer invariant called the spectral flow of the family. showing that our pro blem i s ' elliptic'. this leads to the usual regularity and finiteness results. E) onto the space spanned by the <j>x for A > 0. In fact our problem can easily be fitted into the general class of elliptic problemsf studied for example in (l1). partly because we need the explicit formulae for the heat equation. { E is assumed to have a 0°° Hermitian inner product and Y a (7°° measure dy. The main results of this paper were announced without proof in (5). ATIYAH. The basis of our general index formula (3-10) is in principle similar to that of (2) in that it uses the asymptotics of the heat equation. The main technical and computational part of the paper is in section 2 where we consider our boundary value problem on the semi-infinite cylinder Y x R+ instead of X. by the results of Seeley (16). We now form the product Y x R+ of Y with the half-line u > 0 and consider the operator s DM (2-1) acting on sections/(?/. F.48 M. PATODI AND I. put B = A+H where H denotes projection onto the null-space of A. then B is invertible and so. The computation of this boundary contribution can be made on the cylinder and much of section 2 is concerned with this calculation. \B\ the positive square root of B2 is pseudo-differential. To see this. (2-2) | In the framework of (11) our problem would be an over-determined elliptic system. Let P denote the projection of C°°( Y. E) a self-adjoint elliptic first order differential operator. M.

it will be sufficient to prove (iv) by estimating the Sobolev norms. we expand / and g in terms of an orthonormal basis of eigenfunctions of A: f(y. PROPOSITION (2-5). P) such that (i) DQg = g for all ge C»mp( YxR+. z. To solve Df = g. u) = £ / A ( « ) 4>\(y)> 9(y> u) = ^gx(u) <j>x{y).CcwomP(Tx R+>E) ->C»(7x R+ E. I 49 We now impose the following boundary condition for D P/(-. Also Hl (for I ^ 0) will denote the Sobolev space of sections with derivatives up to order I in L% and H[oo the space of sections which are locally in Hl.Spectral asymmetry and Biemannian geometry.E) (ii) QDf = fforallfeCc»mp(YxK+.. (2-6) AW = Jo / • for for Ju A<0 (2-7) to define Qk. ' (2 8) ' PSP 77 . we rewrite the equations (2-7) in terms of the Laplace transform Jo and we get / ( § ) = gA(g) +/A(0) X + iE.0): in fact it is equivalent to /. Then the following proposition asserts that (2-1) with the boundary condition (2-3) has a good fundamental solution Q.zeY and u. vanishing for u > 0). Proof. (2-4) The space of all C°° sections satisfying (2-3) will be denoted by C<°( Y x R+. v) of Q is C<° for u 4 = v {here y. We must now solve = 0 for We take the explicit solutions A ^ 0. veR+) (iv) Q extends to a continuous map H'-1 -> H[oofor all integral I > 1.P) (iii) The kernel Q{y. 0 ) = 0. (/(y. Jr The adjoint boundary condition to (2-3) is clearly ( l .O). To do this.e. (2-3) Note that this is a global condition for the boundary value f{'. E\ P) and ^comp will denote sections with compact support (i.0) = 0.u.P ) / ( . More precisely <peE} i£B(8/8uy^> eL2 for all j and for all differential operators B on Y of order I —j.0A(y)) = O for all A > 0 .E. [Formally Q = SQA satisfies (i) and (ii) and to see that this formal solution for Q converges. There is a linear operator Q.

Jo 2 Computing Z -norms from (2-8).e ( .9) by powers of A and differentiating (2.50 M. For A = 0. this gives •/A Uitt/ II < . For 0 < t0 < tx < oo. we have Hence.< ) e t u l ( l .6) with respect to u. (2-9) and (2-10) therefore give continuity H° -> H1. the Sobolev norm || || x for H1 may be defined 1 1 != Except for A = 0. I t remains to verify (iii). For t > 0. fo(u) = f" gr0(v)rfuandthis. A T I Y A H . M.t w l P . this can be deduced as follows. P A T O D I A N D I .togetherwith(2-10). multiplying (2. where e{t) is the characteristic function of the non-negative real line and ( 2 "H) \A\ =AP-A(l-P). we deduce for all A (2-9) Combined with (2-6).P ) . K . z. (0) = — eXvgx(v) dv for A < 0.o i l ™ II Since A is a first-order elliptic operator. t) is a C00 function on Y x Y x R+. S I N G E R where/ A (0) = 0 for A > 0 and / .givescontinuityf?0^-ZfJooestablishing Jo (iv) for I = 1. we deduce which establishes (iv) for general I. Alternatively. Now Q is given by convolution in the u-variable with K{t) = e ( « ) e . F . using Parseval's formula. e~*Ul is t h e ' heat operator' associated to the pseudo-differential operator \A | and so its kernel E{y. we have 2 dydzdt = P (S (h-t0) oo . V. More generally.

If we decompose the Hilbert space L2( Y x R+. We shall give the explicit kernels of these operators in terms of the eigenfunctions 0A of A. Now let us form the two self-adjoint operators Ax = S)*Sl.u. Then R' = (Q')* follows by continuity from the fact that (Df. where K(y. D* = — 8/du and the adjointness is clear. Since Q(y. OnH". Thus D* in the above proof occurs for A = 0 in which case we put/0(%) = — o Ju also has a fundamental solution with the properties in (2-5). Applying Proposition (2-12) we then see that equality must actually hold. u) = S/A(w) <f>x{y). = v. (2-14) (2-15) •(• Note that restriction is even continuous fl'jFx R+) ->• Hi(Y). If we regard D and D* as unbounded operators on L2 we have 2 PROPOSITION (2-12). H" involving the zero-eigenvalue of A and H' all the non-zero eigenvalues.D*g) for /. we can then consider the bounded operators e~tAi and e~tA*. Proof. namely Domain Si = W and similarly for Domain Si*. The closure of the operators on L defined by D andD* with domains given by (2-3) and (2-4) respectively are adjoints of each other. Consider first Ax. I 2m 2t |A| 2N 51 (since £A e~ ° < C1iX~ which converges for large N). z. g e C^omp and satisfying (2-3) and (2-4) respectively. Let Si denote the closure of the operator D on L2 with domain given by (2-3). we see that for each A we = 0 with the boundary conditions must study the operator — d2/du2 + A2 on u 5 /A(0) = 0 if if A> 0 A < 0.Spectral asymmetry and Riemannian geometry. (2-13) Expanding in terms of the <f>x. E) into two parts :H = H' ®H". z. k the Sobolev Lemma implies that E is C°°.z. which is the kernel of the composite continuous map #*( Y x R+ E) -^ L\ Y. For t > 0. Since adjoints commute with inverses. Q has a C*1 kernel for u 4 If we replace D by D* and the boundary condition (2-3) by (2-4). the proposition is established. the same is therefore true of K(t) for t # 0. which is the operator given by — d2/duz + A2 with the boundary condition = 0 and (1-P)1(?L + AA j = 0. By (2-11). so that/(y.u — v). t) denotes the kernel of K(t). v) = K{y. On H' the fundamental solution Q of (2-5) gives a bounded inverse Q' for D and similarly we get a bounded inverse R' for D*. E) •£ L\ Y. the only difference f °° g (v) dv. then D and D* decompose accordingly.f A similar remark applies with D replaced by D* and (2-3) replaced by (2-4). 4-2 . A2 = S>S>*. where r is restriction to the boundary. The domain of Q> is clearly contained in the closed subspace W. D = d/du. Since this holds for all j . E).g) = (f.

More precisely they are bounded by PROPOSITION 0 for some constant G as t->0.{y) <j>\{z) and summing over all A. V. section 4)) we deduce (2-21). A T I Y A H . The kernels ofe~mi and e~tA* are exponentially small intast-* for u = ) = v. standard Laplace transform methods (see ((6). it follows that the kernel ex(t.u.v) of e~tAi is bounded by (2-20) Since the kernel of e~tA* on the diagonal of Y x Yis bounded by Gt~nl2 (see for example ((2). as usual. section 14-2)) give where erfc is the complementary error function denned by erfc (a. K . P A T O D I AND I. For the operator A2. Moreover we are interested in the difference between . «=o (2-18) (2-19) The fundamental solution of d/dt — d2/du2 + A2 for (2-18) is again given by (2-16) while for (2-19) we must use (2-17) with A changed to — A. y. Proposition (2-21) asserts that. the contribution outside the diagonal (u = v) is asymptotically negligible. the boundary conditions for each A are A(0) = 0 if A < 0 = 0 if A Ss 0.) =-p-1 The kernel of e~tAi at a point (y.v) is therefore given by multiplying (2-16) or (2-17) (according as A ^ 0 or A < 0) by <j>x. Since e-^dg < e~x2 we see that (2-16) and (2-17) are both bounded by Using the inequality x ^ ex^2.u\ z. What we are primarily interested in is of course the contribution from the diagonal. M. F .52 M. z. S I N G E R Now the fundamental solution for 8/dt — d2jdu2 + A2 with the boundary condition (2-14) is easily seen to be while for the boundary condition (2-15).

so let K(t. rj(s) is holomorphic near 8 = 0 and its value at s = 0 is given by (2-27) Finally. Integrating over Y x R+. as t ->• oo in (2-23). we have denned sign A = + 1 if A = 0. y. (2-17) and their counterparts for A2 then give K{t. where dN{s) is holomorphic for i?e(s) > . K(t) + \h ->. for convenience. Integrating by parts and using (2-24). we obtain K(t) = f" f K{t. S] for some S > 0. In particular if we assume that K(t) has an asymptotic expansion K{t) ~ then (2-25) yields £ aktik as * -• 0. we integrate only over Y x [0. differentiating with respect to t. In particular. This is the final formula which we shall be applying in the next section. where h = dimKer^4 is the multiplicity of the 0-eigenvalue. (2-23) Note that. instead of defining K{t) by integrating over Y x R+. Formulae (2-16). Also (2-23) shows that \K{t)\ < 1/V77-Se-A2t < C7«-4" as < ^ 0.y.Spectral asymmetry and Riemannian geometry.u) = j ^ { ^| ^ | ^ ^] )J (2-22) where. I tA tA tA tA 53 e~ i and e~ *.y.y. . we note that if. In fact.£"(<) -> — \h. u. u) denote the kernel of e~ i — e~ * at the point (y. Moreover.\{N +1). Thus (2-26) gives the analytic continuation of 7](2s) to the whole s-plane.u)dydu = _ S ^ ^ e r f c ( | A | V<) Jo J Y K * and hence. the asymptotic expansion is unchanged.0 exponentially as t -> oo.u) of (FxR+) x ( 7 x R + ) . Hence for i?e(«) large Jo o converges. the difference is S sign A^e2^:* erfc (S/Jt + \ A| Jt) A . we get (2-25) by definition of rj(s).

More precisely. Let D: C°°(Z. such that p = 0 for % < a . a = crD(du) is the bundle isomorphism E -> F given by the symbol aD of D and Au: OGO( Y. E) ->. Finally. We consider the operator D with the boundary condition (2-3) and we construct a parametrix R by patching together the fundamental solution Q of (2-5) with an interior parametrix Q2. With this assumption.and p = 1 for u ^ b. The index formula.Cc0( Y. we assume A is self-adjoint. i/r2 being extended by 1. F) be a first order elliptic differential operator. We now assume that Au is independent oiu. F . and define four (700 functions <f>x. Eu) is an elliptic operator on Y depending on the parameter u (and Eu = E\ Y x {«}). V. we shall just write our operator in the form As in section 2. In particular. K . More precisely. b) denote an increasing C00 function of the real variable u. it is then of the form where us I is the normal coordinate. We regard these functions of u as functions on the cylinder Y xl and then extend them to X in the obvious way: 0 1. 0 a . The main result is stated in Theorem (3-10). \jrx being extended by 0 and <f>2. 3. we put Qx = Qa—1 and (considering (j)it i/ri as multiplication operators) . Eu) -s. <r is taken to be an isometry. we shall consider a boundary value problem on a compact manifold X with boundary Y which coincides near the boundary with the problem studied in section 2. The computations in section 2 will then be used to derive an explicit formula for the index of our problem. F extending this inner product from Y x I to X and a Cm measure dx on X extending dydu on the collar. In a neighbourhood Y x I of the boundary.54 M. In this section. M. let p(a.C°°(X. A T I Y A H . xjr^ i/r2 by Note that ^ = 1 on the support of ^rt. S I N G E R which is bounded by and hence is exponentially small. with respect to given C00 hermitian inner products on E\ Y and a C°° measure dy on Y. and we shall fix inner products on E. P A T O D I AND I . this means that a suitable isomorphism of E £ n*(E0) identifies all Au with A = Ao (here n denotes the projection Yxl ->• Y).

Switching the roles of 0 i5 fa gives a left parametrix. (3-4) JOJ T f Note that this is not the e2 of section 2 which was the counterpart of e t but for the operator . E. and indexZ> = d i m K e r ^ . orthogonal to the image of D. From this. Similar statements hold for the operator D* with the adjoint boundary condition (2-4) and by essentially the same argument as in (2-12). for t > 0. Proposition (2-5) also shows that R is continuous from H1-1 -> Hl (for I > 1). that is DR — 1 has a (/"-kernel. The operator &I*3I is then a self-adjoint operator and e~ts'B is the fundamental solution of 8/dt + D*D with the boundary condition (2-13). l where * denotes convolution in t and composition of operators. there is a natural double of D on the double of X.Spectral asymmetry and Riemannian geometry.> J as t -*• 0. P) is the space of sections of E which satisfy the boundary condition (2-3). F) (for 1 ^ 1 ) are Fredholm operators with the same null-space. we put f=<l>1e1rjr1 + 4>2e2f2. (3-2) where Q> is the closed operator defined by D (with domain H^X.m = cm_x*c1. P) -> Hl~i(X. be taken as the Green's operator for D on the double of X (note that.d i m K e r ^ * . u) fa(u) dydu+[ JX F(t. and the fundamental solution1)e2 of d/dt + D*D on the double of X.fadefined above.: H \ X . is necessarily C^ (being in Ker-Df). Thus.E. a 2-sided parametrix. e has a C°° kernel which differs from the kernel of / by an exponentially small term as t -*• 0. it follows that this gives the Hilbert space adjoint of D (as a closed operator on Z2). I 55 Here Q2 can. the roles of E and F being switched on the two halves). An approximate fundamental solution can be constructed from the fundamental solution elt constructed for the corresponding operator on the cylinder in section 2.F) Z>. This implies that an _L2-section of F. E. it follows (see for example(l3)) that our fundamental solution e = e~ts'B is given by a convergent series of the form e=/+ S (-l)mcm*f. It now follows that and D:CX(X. in fact. computed either in C00 or in L2. Using the functions <f>t. in computing the asymptotic behaviour of Tr e~ts'B. x) fa(x) dx. Proposition (2-5) shows that R is a right parametrix. for definiteness. R is a linear operator where C™(X. Here cx = (d/dt + D*D)f and c. (3-3) Proposition (2-21) and the corresponding result for closed manifolds show that (d\dt + D*D)f is exponentially small as t-> 0. and the fact t h a t / .E. y. we can replace e by/. In particular. Applying the same remarks to e~t3B* and subtracting.P)^C*>(X. we obtain asymptotically ®0' ~T [ K(t. hence R is. P)). because of the form (3-1). Thus D has a well-defined index.

Trc-*»»* ~K(t)+ £ f ak(x)dxtik. we know (see ((2).(*) Is (i runs over the eigenvalues of D*D (for i = 1) or DD* (for i = 2) and 0A are the corresponding eigenfunctions. x) ~ 0 for x in the collar. this gives an asymptotic expansion for K(t): K{t)~iadexDApplying (2-26) we deduce that £ f ak{z)dxtlk. F(t. x) ^rz{x) ~ F(t. x) and so (3-4) becomes *9 . K . Also the null-space of Si*S> coincides with that of 2. via <r. Because of the estimates at the end of section 2. In the collar neighbourhood of Y. the operators DD* and D*D are isomorphic. On the closed manifold (double of X). x)—F2(t. (3-5) where the coefficients ak(x) are explicit local functions of the operators DD* and D*D. (3-7) k>-njX 2sJn ^ where = ^N(S) is holomorphic for Be(s) > iV+1 —.56 M. V. More precisely. x) = Fx{t. S I N G E R where K is defined by (2-22) and F(t. in the first integral by so that we end up with the function K{t) o Jo given by (2-23). F . then 2)2)*{3)4>) = /iQi<f> so <f>\-^!3<f> defines an isomorphism of the/i-eigenspace of S)*2> onto that of S>S>* (with inverse i/n-^/i-1^)*^). if p4. A T I Y A H . (3-6) Now the operators ^ * ^ a n d ^ ^ * have discrete spectrum with finite multiplicities and their non-zero eigenvalues coincide. In fact. where 10. section 4)) that we have an asymptotic expansion F{t. z) is obtained from the kernels o£e-tD*D and e-tDD' on the double of X. M.x) ~ 2 <xk(x)ti«. Hence. and so F(t. it follows that. Hence by (3-2) index D = Txe-t3>'3Combined with (3-6). (3>8) Thus 7]{s) is extended meromorphically to the whole plane and in particular ?/(0) = 2 f ao(x) dx-(h + 2 index D) Jx or index D = I ao(x)dx tr~^(3'9) z . while the null-space of 3)3)* coincides with that oiQ*. x). P A T O D I AND I. F(t.

^ denote the eigenvalues and eigenfunctions of D*D on the double of X. <j>" are the corresponding objects for DD*. Let X be a compact manifold with boundary Y and let be a linear first order elliptic differential operator on X (acting from the vector bundle E to the vector bundle F). one has to work somewhat harder and this will be discussed in Part III. Moreover. the natural operators A arising in Riemannian geometry tend to be extendable in this sense. -co)} illustrated in the figure. and /*". A*0 In (iii) the series converges absolutely for Re(s) large and then r/(s) extends to a meromorphic function on the whole s-plane with a finite value at s = 0. 7/ are defined as follows: (i) ao(x) is the constant term in the asymptotic 'expansion (as t -»• 0) of = fJ x where ft'.E. The index of the boundary value problem in Theorem (3-10) can be given an alternative description by introducing the non-compact manifold l = Z u { 7 x [ 0 . (ii) h = dim Ker A = multiplicity of ^-eigenvalue of A (iii) 1/(5)= 2 sign A | A|~s. if the asymptotic expansion in (i) has no negative power of t then v(s) is holomorphic for Re(s) > — \. I 57 This formula is what we have been aiming at and we therefore summarize our results so far in the following theorem: THEOREM (3-10). D takes the special form where u is the inward normal coordinate. a is a bundle isomorphism E\ Y -> F\ Y and A is a self-adjoint elliptic operator on Y.F) has a finite index given by indexZ> Jx where a0. Let C°(X. Remark. . The proof depends on being able tofindX with dX = Y and an elliptic operator DonX extending 8/du + A. associated to the operator A on Y. We assume that. where A runs over the eigenvalues of A. Then D: C«>(X. As we shall see in the next section. in particular that v(0) is finite. h. P) -> C°>(X. E. This theorem can be viewed as providing information about the function v(s). For more general operators.Spectral asymmetry and Riemannian geometry. in a neighbourhood Yxl of the boundary. P) denote the space of sections f of E satisfying the boundary condition where P is the spectral projection of A corresponding to eigenvalues > 0.

Thus/has fx as an asymptotic or limiting value as u -> — oo. Since Df = 0. F. Then (i) K e r D is isomorphic to the space of L2-solutions of Df = 0 on X. (i) Let fe Ker D and expand it near the boundary in the form where the <j>x are an orthonormal base of eigenfunctions of A. (3-11). We shall be primarily interested in the i 2 -sections of E and F over X. hence certainly in 2A Conversely a solution of Df = 0 on X. because terms involving e~Ku with A ^ 0 are not in L2. — oo).F in the cylinder Y x [0. By an extended 2/2-section of E.58 M. On the compact manifold X we have already seen that Ker D coincides with Ker D*D (with the appropriate boundary conditions). we shall mean a section/which is locally in L2 and such that. Thus/)->•/ gives the required isomorphism. must be of the form (3-12). F (together with their inner products) to X in the obvious way and we take the measure dx = dydu on Y x [0. where g is in L2 and / „ e Ker A. We shall now show that similar results hold for Zr2-sections on X. — oo)).sothat/A(M) = e~A«/A(0). but we shall also consider a slightly larger space. SINGER We extend the bundles E. then instead of (3-12). PROPOSITION X Proof. (3-12) A<0 This shows that/extends to asection/on X which satisfies Df = 0 and is exponentially decaying as u ->• — oo. for large negative u. K. an extended Z*2-section/of F satisfying D*f = 0 is necessarily of the form (3-13) so that/(->/gives an isomorphism as required. h(F) the corresponding dimension for D* and h^F) is the dimension of the subspace of KerA consisting of limiting values of extended L2-sections f of F satisfying D*f = 0. PATODI AND I. COROLLARY 2 . Since Pf(. (ii) KerD* is isomorphic to the space of extended L2-solutions of D*f = 0 on X. we have df/du + Af = 0. ATIYAH. V. E) the adjoint operator. M.u)= Z AS=O a n d / i s now an extended £2-section of F (as defined above) with limiting value y)(3-13) Conversely. A similar definition holds for extended Z2-sections of F (using the isomorphism o~: E -». 0) = 0. we have/ A (0) = 0 for A ^ 0 and so /(p)=Se-kA(0)^).E. (ii) Let/eKerZ)*. we get f(y.l-P)^C*>(X.P)^Cco(X>F) be the operator in (3-10) and D*: C (X. which is in L2. In view of (3-2) we therefore deduce (3-14) index D = h(E)-h(F)-hoo(F) where h(E) is the dimension of the space ofL solutions of Df = OonX. F. Let D: C'a(X.

af)dy. it follows that the right-hand side of (3-18) tends to zero as U -> — oo.oo) of the form f(y. (3-21) where hm(E) is the dimension of the subspace of Ker^4 consisting of limiting values of . / i s exponentially decaying as u ->• — oo. . applying Corollary (3-14). the dimensions in (3-14) can equally well be defined using D*D and DD*. we have |/('. Theresolutions ofD and D*D coincide on X. we must have aA = 0 for A < 0 and bx = 0 for > 0. As before. we see that Df also satisfies an estimate of the form (3-17). We now apply Green's formula to the compact manifold Xv = I u [ 0 .Df)dx=0 (3-19) and hence Df = 0. P) ->• C^iXjE). as do the extended L2-solutions. This proves the first part of the Proposition. We conclude this section with some further comments on the 0-eigenvalue. The same also holds with D and D* interchanged. Hence. I 59 PROPOSITION (3-15). Applying D to (3-16). (3-17) where I || denotes the Z2-norm on Y. P ] c l obtaining f J Xp (D*Df.tt)|| is bounded as u ->• —oo. (3-18) gives f (Df. Since D*Df = 0. 1 -P) -> C°°(Z. where P is essentially the spectral projection of A corresponding to A < 0. E) and hence a different index. Any solution o£D*Df = 0 has an expansion on Y x [0. F. However applying D removes these terms and so (3* 17) holds for Z)/and this is now enough to show that the right-hand side of (3-18) tends to zero as U -> — oo. In fact. For an extended ^-section/ satisfying D*Df — 0. In view of this Proposition. proving the second part of the Proposition. the expansion (3-16) may now have non-zero terms corresponding to A = 0 so. we can only assert that ||/(*.u) = "Ei(axeXu + bxe~Xu)^x(y) A (3-16) and for this to be in L2 as u -*• — oo. we obtain an operator^: C^iXjF. we conclude that Df = 0.f)dx-( J XJJ (Df.Spectral asymmetry and Biemannian geometry.Df)dx=( J YJJ (Df. The corresponding statements with D and D* interchanged are proved in exactly the same manner. we get index B = h(F)-h(E)-htc(E). instead of (3-17).Inviewoftheestimates(3-17)for/andZ)/. If we replace!) by D* from the beginning. More precisely. (3-18) where J ^ = dXv is the copy of Y at u = f7.w)| < Ce"tt with a > 0. This is because near the boundary of X we have Thus # h a s a slightly larger domain than D*: C°°(X. Proof.

h(E) . Then the Dirac operator of X is an elliptic first order differential operator D:Ca>{X. Finally.hx(E) l (3-23) while Theorem (3-10) and Corollary (3-14) applied to D give J' \ao(x) dx - + ^( ) = h(E) .h(F) .hw(F). A T I Y A H . On the other hand. that the integrand oco(x) dx that occurs in Theorem (3-10) can be identified explicitly as a suitable Pontrjagin form and that the asymptotic expansion in (3-10) (i) has no negative powers of t (implying. then. section 6) or ((12). according to Theorem (3-10). Theorem (3-10) applied to I> gives indexZ)= — ao{x)dx 7T~^. S I N G E R extended Z/ -sections f of E satisfying Df = 0. and in 7 x / w e have where A is the Dirac operator on Y. Note that the Dirac operator on Y depends . (3-22) * Jx where a0. F . The prototype. that X is a spin manifold and we choose a definite spin structure. The restriction of S+ to Y may be identified with the spin bundle S of Y (associated to the spin representation of Spin (2n — 1)). 4. K. h. the dimensions behave as if this were true and we shall use (3-25) in the next section for one of our applications. This formula suggests that every section in Ker A is uniquely expressible as a sum of limiting values coming from E and F respectively. (3-24) Adding (3-23) and (3-24) we deduce (3-25) h = h<a{E)+hx{F).8+)-+C*>{X. that v(s) will be holomorphic for Re (s) > — £). We assume.8-). In any case. M. Let X be an oriented compact manifold with boundary of dimension 2n.60 2 M. and in some sense the fundamental case. For its definition and formal properties we refer to ((2). We shall now apply the general index formula of section 3 to the cases that naturally arise in Riemannian geometry. The Riemannian case. moreover. The main result of (2) asserts. Thus (3-21) and (3-22) yield ao(x) dx Jx = 5 ^ = h(F) . Chapter IV). where S+ and S~ are the two spin bundles (associated to the two half-spin representations of Spin (2n)). u is the inward normal coordinate and <r is Clifford multiplication by the unit inward normal. P A T O D I AND I. we choose a Riemannian metric on X which coincides with a product metric o n T x / i n a neighbourhood of the boundary. 9/(0) refer to the operator D (the point being that ao and 7/(0) change sign when we replace D by 3 but h = dimKer A = dimKer( — A) is unaltered). so that Y = dX has dimension 2n — 1. V. is that of the Dirac operator and so we begin by describing this.

the case Y = circle.2). dimX = 2 mod 4. a solution of Af = 0 (or equivalently A2f = 0) is called a harmonic spinor on Y. Hence Spec A is symmetric about the origin. v(s) is the v-function of the Dirac operator on Y and A(p) is the Hirzebruch A-polynomial applied to the Pontrjagin forms pi of the Biemannian metric on X. these integers being computed from the non-compact 'elongation' X of X as in section 3. This was known before but only as a corollary of the mod 2 index theorem ((4). Now h mod 2 is directly seen to be an invariant of Y. On the other hand the integral in (4-2) remains unaltered. We now distinguish the two cases n = 1 and K E 3 mod 4. (a) n s 1 mod 4. then A(p) = 0. On the one hand it is fairly clear that the analytical results of section 3. where B is real skew-adjoint (cf. I 61 on the choice of orientation (it changes sign if we reverse the orientation). using the particular structure of the Clifford algebra in these dimensions (see (3)). i. so n(s) = 0 and (4-2) reduces to: h = —2 index D. Together with (3-25) this yields the formula given in Theorem 4 of (5). independent of the metric (see (4)) and so (4-2) gives an analytic proof that this invariant vanishes when Y is a spin boundary. provided the first two normal derivatives of the metric vanish on Y: this is because two such metrics yield a C2-metric on the double of X. Remarks (1) Using (3-14).sodim Y = 5 mod 8. Theorem (3-1)). Moreover TJ(S) is holomorphic for Re(s) > — \. The sign convention we are using may be considered as defined by (4-1). and so we can argue as in section 1. or in other words A mod 2 is a spin-cobordism invariant.Spectral asymmetry and Biemannian geometry. For a more general metric. the spin-bundle of Y has a natural quaternionic structure (defined say by j) and jA = — Aj. Combining Theorem (3-10) with the results of ((2). concerning operators like D. As usual. (4-2) continues to hold as it stands for such metrics. Thus. This shows again that n(s) = 0 and also that the harmonic spinors have a quaternion structure. then the Dirac operator A on Y is of the form A = iB. . continue to hold: the fundamental solution given in section 2 must now be used as simply the first step in an iterative procedure to construct a parametrix. (3) If n is odd. Thus index D will be defined and will remain constant under continuous variation of the interior metric (while the metric on Y remains fixed). so dim Y = I mod 8. and this is the analogue of the geodesic curvature in Gauss-Bonnet. (b) n = 3mod4.e. we obtain (4. In this case. Thus h is automatically even and (4-2) gives no further information (except to identify \h with — index D). there will be an extra integral over Y involving the second fundamental form. section 6). The index of the Dirac operator on X with the global boundary condition (2-3) is given by h + V(0) index D = A(p) — THEOREM -L- where h is the dimension of the space of harmonic spinors on Y. (2) The condition that X be isometric to a product near Y can clearly be relaxed. we can replace index D above by h(S+) — h(S~) — hx(S~). B = d/dx).

. the metric and connexion are constant in the normal direction. F. On the other hand. x 1 = 2-2Sr-r. This identifies the elongated manifold X conformally with the punctured surface X— U pi.. V. (4-5) I t is interesting to check (4-5) directly using (3-14). this includes the case of complex Kahler manifolds. we have h(E) = 0 (assuming r > 0.pr. whose boundary Y consists of r circles. the space of holomorphic £2-one-forms is a conformal invariant (the condition J w A w < oo does not use the metric) and hence can be computed from the punctured surface X — \J pt. Applying now the analogue of (4-3) for the 8-operator on X. by Gauss-Bonnet. As explained in ((2). where g is the genus of X Thus (4-4) becomes index 8 = 1— g — r. (4-4) Here we have used the fact that the tangential component of d on each boundary circle is the standard operator —1(8/86). (4-2) generalizes to f ^tai^ (4. J. Now the Chern form cx of a surface coincides with the Euler form and so.62 M. sJ is the sum of all Hirzebruch polynomials Ak and we pick out the form in the product of top dimension. defining a spinmanifold being now relaxed to allow o)2{X) to be the image of an integral cohomology class). If zt is a local parameter centred at pit we choose a metric on X— \Jp t. One easily checks that a holomorphic L2 form must then .. Then X = XU Bt i= l is a surface.. As a simple illustration. Using the results of ((2).. which is | dz^ | 2 near p t . acting on spinors with coefficients in £ (that is. K. let us consider the case of Riemann surfaces.BT around points p v . so that X + X).3) where hg. ijg(s) relate to the operator Ag and in the integrand ch £ denotes the Chern character of £ (as a differential form). PATODI AND I. Let X be a compact Riemann surface (without boundary) and delete from it small disjoint open discs Blt . SINGER Suppose now £ is a hermitian vector bundle on X with a unitary connexion and that. M. we tensor the spin bundles with E) and near Y we have where Ag is the generalized Dirac operator on Y. Similar formulae hold for spinc-manifolds (the condition (02{X) = 0. Since a holomorphic Z2-function on X decays exponentially at oo. we obtain Jjj-jl. Then we have a generalized Dirac operator Dg. section 6). near the boundary.. hence rj(s) = OandA = 1. section 6). ATIYAH.

the ^/-function of B is twice the ^/-function of 5 e v . one checks easily. Note that B preserves the parity of forms on Y and commutes with <j> i-> ( . and hence defines by restriction an operator which we call the signature operator. Combining Theorem (3-10) with the local signature theorem of ((2). where sign (X) denotes the signature of the quadratic form on H2k(X. and L is the Hirzebruch Z-polynomial. A is of the required form with B a self-adjoint operator on Y. and the same holds for the dimension of the null spaces. Moreover. we can replace index A by indexA = h+-h~ — h^. (4-7) Jx where h and T\ refer to the operator Ber. as in (4-2) and (4-3). To compute h^F) we use (3-25) and the fact that h^E) = 1 (extended holomorphic L2 functions on X being just the constants). Included amongst the case of (4-3) is that of the signature operator. section 5) we deduce. In fact we may identify the restriction of Cl+ to Y with the space of all differential forms on Y and a little computation then shows that B<j> = (-l)k+P+1(e*d-d*)<fi. so that h(F) = g. This is of particular importance because of its relation with cohomology and so we shall now treat this in detail. In particular. R) given by the cup-product. (4-8) where h± are the dimensions of Inharmonic forms in Q ± of the elongated manifold X and ft~ is the dimension of the space of limiting values of extended L2 harmonic forms . (4-6) where * now denotes the duality operator on Y. of all differential forms and anti-commutes with the involution T defined by T<j> — i»(p-i)+' * ^ for <j>e£lv. the operator d + d* acts on the space Q. Applying (3-14) this gives index 3 = 0 — g— (r— 1). Denoting by Q+ and fi_ the + 1-eigenspaces of T.\)v * <j). it follows that d + d* interchanges Q + and £l_. 4k = dim X and <j> is either a 2^-form (e = 1) or a (2p — 1) -form (e = — 1). for an oriented Riemannian manifold X of dimension 21.Spectral asymmetry and Riemannian geometry. so that B = Bev © BoAi and Bev is isomorphic to Boid. Also by (3-14). Suppose now X is a manifold with boundary Y and is isometric to a product near the boundary. For a closed manifold X of dimension 4k. indexA=\ L(p)-(h + 7](0)). We recall that. I 63 extend holomorphically across the puncture. that index A = sign(X). we need to give it an independent treatment. since the signature operator exists without the spin restriction involved in (4-3). which checks with (4-5). then near Y. by the Hodge theory.

6 = do + dxdu and Yu = dXv is a copy of Y. 169) asserts that any closed Z2-form i/r can be written as i/r = dd + 0. To see that [0] lies in the image of H*omp(X). Since the Laplace operator preserves degrees of forms the components 0 9 of a harmonic form 0 = E0 9 are also harmonic. In particular. ATIYAH. The space 3f{X) of L2harmonic forms on X is naturally isomorphic to the image &{X) o / # * o m p ( l ) -> H*(t) {or equivalently H*{X. F. (4-9) shows that a(0) = £. Y) -+ H*(X)). In particular. depending on u. that 6 can be chosen bounded (i. First. it follows that 0 extends to a C* form on X Hence 0 = d# with 0 a C^-form on X Since dv = — dufu2. We proceed now to give topological interpretations to the integers occurring in (4-8). Since 0 is exponentially decaying as v -> 0. f 0 = f 0> J y J ya where y is a cycle in Y and yu the corresponding cycle in Yu = Y x {u} c: X. a closed C°° form i/r with compact support has such a decomposition. where 60. Since d*<f> = 0 and </> = dd. PATODI AND I. p. Now. 0 defines an element [0] of H*{X). Finally. In fact. it is sufficient to check that its restriction to H*{ Y) is zero.64 M. A quite general theorem of de Rham-Kodaira ((9). K. SINGER in Q_(X). to prove injectivity we assume a(0) = 0. we deal with surjectivity. dx are forms on Y.e. V. putting v = \ju we get a normal coordinate near the boundary for the compactification X of X (obtained by adjoining a copy of Y' at oo'). d<f> = d*<fi = 0 and 6 is some current (distributional form). the boundary contribution in (4-10) tends to zero. We claim first. 6 = 60 + dxdv. PROPOSITION Proof. since <j> is closed.. and we get f J (4-10) where <j> = fo + fadu. it follows that 60 and 6xu2 are both bounded. But 0 is exponentially decreasing as u ->• — oo (see (3-17)) hence Jy 0 = 0 for any y and so [0] restricts to zero in H*{ Y). Because of (3 • 15) the L2 harmonic forms on X coincide with the L2 solutions of {d + d*) 0 = 0. Since every £ei?(X) is represented by such a ft and since H*{X) can be computed from the complex of currents. we deduce that . (4'9) where 0 e i 2 . so that 0 = dd for some C™ form 6 on X. First we prove the following result which is of some independent interest: (4-9). Now apply Green's formula to a compact part XJJ of X (given by — u ^ — U). and are bounded). as «-> — oo. It follows that an L2 harmonic form 0 also satisfies d<fi = d*(f> = 0. Since </> -> 0 exponentially while 0 is bounded as U -> — oo. M. Thus 0 1 — > [0] defines a map which we shall prove is an isomorphism. so that a is surjective.

if ^±: X'±-*-H*(LX) are the restrictions of /?. Sign (X) — h+-7i~. But 65 f hence <f) = 0 and a is injective as required.±(X). it is pointed out that the Hodge theorem. This quadratic form is induced by the degenerate quadratic form on H2k(X. any ^eJTsatisfies^ = d*<fi = 0 and hence we obtain a map /?: Jf-> H*(X) extending a. To deal with the third term h~. Hence 5 PSP77 . In (9). we can write (j) = \jr + 6. for <j>e Jf±. where A± = d i m ^ * and n Q. just as for a compact manifold. we have xjrxdu= ± T(I/T0) and so xjr0 = 0 ^xpeJt^. We recall now that the signature of a 4&-dimensional manifold X with boundary Y is defined to be the signature of the (non-degenerate) quadratic form denned on S2k(X). (4-9) implies COROLLARY (4-11). Again. we see that. Remark. as defined in section 3. Y) given by the cup-product: Poincare duality for (X. But (3-25) asserts that h++h~ =2h (recall that the h in (3-25) is 6imH*(Y) which is 2A of the present section). does not extend to i 2 -forms on a noncompact manifold. Now Poincar6 duality in the exact sequence H*(X) 4 H*(Y) 4 H*(X. where h± = Him Jf±/j^>± (4-13) in the notation of section 3 and (4-8). ip'1 harmonic forms on Y (independent of u) and 6 an L2 harmonic form (hence decaying exponentially). Y) shows that Imj isjdual to its orthogonal complement. The composition is thus given by <j>\-^^r0. Y) -»• H2k(X) . I as U -> — oo. so that dimlmj = idim#*(F) = h (where h is the same as in (4-7)). Now if JT* = Ct n Q ± . we consider the space Jf = > J^(X) of all extended L2 harmonic forms on X. In the cylinder. there is still an interesting connexion between L2 harmonic forms and cohomology. identifying cohomology with harmonic forms on a compact manifold. Y) shows that the radical is precisely the kernel of H2k(X.Spectral asymmetry and Eiemannian geometry.Thus. we have (4-12) Kerj/?± = ^f±. because of (3-15). for the special cylinderlike manifolds X. This corollary identifies two'of the three terms in formula (4-8). where ijr = t/rg + ^du with \Jf0. Proposition (4-9) shows nevertheless that. From (4-12) we therefore deduce h£ «S h.[Hence.

even/odd).. PATODI AND I. Y) in Hik(X). Now let us replace A by .pk). Then signX = f L(p). A—. and so the integers h. d = £{sign X + E(Kevj)}. Using (4-11) we therefore derive our final result. V. we mentioned the Gauss-Bonnet theorem as an analogue and motivation for the signature theorem we have just arrived at. Using the results of ((2). where Lk is the Icth Hirzebruch L-polynomial and thepi are the Pontrjagin forms of the Riemannian metric. Let X be a ^k-dimensional compact oriented Riemannian manifold with boundary Y and assume that. just Bev whose •(/-function is half that of B: thus 7/(0) in (4-14) gets halved. It is remarkable that the three terms in Theorem (4-14) arise from three different areas: sign X is a topological invariant.. dd -»• Q i T . In section 1. near Y. we must return to the general formula given in (3-10) together with the i 2 interpretation of the index given in (3-14).4+in the proof of (4-14). then.dd. (ii) L(p) = L^Pi. the integral in (4-14) now becomes -{ Z L(p)+ JX e}. and extended L2. We begin by recalling (see ((2).66 M. F.wheree ) is the Euler form of the metric (generalizing (27T)-1 x Gauss curvature for surfaces). SINGER ^« = A» = h. (iii) n(s) is the n-function {defined by (1*7)) for the self-adjoint operator on even forms on Y given by j and is holomorphic for Re(s) > — J. Using (4-9). K. so that A = A+ © A~. . This interplay gives the theorem special interest and leads one to expect a variety of applications. section 6). we deduce dim 3f+v .dim J^-. ATIYAH. We shall now explain this in more detail. it is isometric to a product.. h~ cancel when we combine (4-7) with (4-8). Q^. The tangential operator on y is. where e. However. Tofindthe integer contribution. M. JL(p) is differential-geometric and v(0) is a spectral invariant. From these. we see that sign X in (4-14) must be replaced by (4-15) where Hev( Y) is the even-dimensional cohomology of Y and Stf. namely that in Gauss-Bonnet there is no term corresponding to v(0) in (4-14). the signature theorem for manifolds with boundary: THEOREM (4-14). we also pointed out the significant difference between the two cases. Jf refer to the L2.9/(0). harmonic forms on X of the appropriate types ( ± . section 5)) that the signature operator A: Q+ -»• O_ interchanges even forms and odd forms. Jx where these terms have the following meaning (i) Sign X is the signature of the non-degenerate quadratic form defined by the cupproduct on the image ofH2k(X. (4-16) \JX .

we derive Gauss-Bonnet as the difference between two index problems in which the non-local boundary contribution ?/(0) cancels out. Now we have d*: dO?*1-1 -> dQ 4 *. The cohomological term sign (X) in Theorem (4-14) only involves the middle dimension H2k(X).Spectral asymmetry and Riemannian geometry. Hence we can decompose B in the form B = ev JDQ @ ZJ-L © X>2. Moreover. Bx = d* on dQ. in the course of proving (4-14). Using (4-16). (4-18) (4-17) Finally. (4-15) becomes i{sign X + E(Kerj)} + \$E(Lmj). I ev odd 67 where E = dim H . Adding the two formulae gives (4-14) back again. but subtracting them gives the Gauss-Bonnet theorem E(X) = f e. the Euler characteristic of X. since E(Kerj) + E(Imj) = E(X).2 ". and 2p-l 2p-l *d: p =k = 4Jc-2p— 1 onlyif # 4&-2p + l.1 . Jx From this point of view. (4-17) and the equality dim# C T (r) = £ dim H*(Y) = dim(Imj) proved earlier. To see this.2k~1 and B2 is an operator of the form /O T\ \T* 0 )' 5-2 . Thus dim Jf-dd . we showed that (4-13) was an equality and hence is an isomorphism. where Bo is the zero operator on H . In fact the analytical invariant 9/(0) can also be shown to involve only the 24-forms on Y.ev and observe that the operator B of (4-6) annihilates HeT and coincides up to sign with d* on dQ odd and with *d on d*Q. we deduce the refinement of (4-14) which we want: ±( (4-19) Replacing A+ by A~ yields the same formula except that E(X) and e are changed in sign.oii. In fact.dim # and j is the restriction H*(X) -* H*(Y).dim JT ~dd = dim (Im j)° dd . we use the decomposition Qev = flev @ dQ odd © d*Q. the topological obstruction mentioned in section 1 to the existence of a classical local boundary condition for the operator d + d*: Qev -> Qodd vanishes and one can derive Gauss-Bonnet from the index theorem as given in (l) in which there is no non-local contribution.

If da is an eigenvector of d* with eigenvalue A. we need a metric and the value of the signature will indeed depend on the metric. in odd dimensions. then Q(a) = (a.68 M. M. and SHAPIRO. R. Thus — Q(ct) has the same sign as A. SINGER Since B2 is conjugate to the operator 0 and since ^(TT*) and J(T*T) have the same spectrum it follows that the spectrum of _B2 is symmetric so that its ^-function vanishes identically. Clifford modules. The interest of this reformulation is that (4-21) does not involve the metric. The index theorem for manifolds with boundary. BOTT.d*da.3-38. M. 279-330. . (Oxford 1964). This Proposition leads to a suggestive way of looking at 9/(0). everything bounds after multiplication by some integer. V. BOTT. Similar remarks apply to the ^-functions of the Dirac operator and its generalizations as in (4-3). ?/(0) is finite. In Part III. 19 (1973).. etc. On the heat equation and the index theorem. Differential Analysis (Bombay Colloquium). V. M. F. and we shall then deduce thefinitenessfor all self-adjoint elliptic operators on odd dimensional manifolds. *da) = A-1 (a. However to give proper meaning to its signature. (2) ATIYAH. Thus we have established PROPOSITION (4-20). Invent. Topology 3 (Suppl. R. *d*da) = = -A~ 1 (da. -\-\a. da). K.) Q(a) =[ a Ada. REFERENCES (1) ATIYAH. for the operator on Y defined in (4-14) (iii). Clearly the ^-function on 2 Y is twice the ^/-function of Y and so the finiteness at s = 0 follows from Theorem (4-14). M. This assumes however that we can find an oriented manifold X with boundary Y. Hence the 77-function of B is equal to the rj-function of Bx. F . we can always find an X whose boundary consists of two copies of Y (with correct orientation).. so that Q can be viewed as a non-degenerate quadratic form on dQ2fc~1 (note the formal analogy with the quadratic form leading to sign X). A. In each case we need to use the appropriate cobordism theory (see (17)) to deduce that. We conclude with a few remarks about the finiteness of 77(0). A T I Y A H . and PATODI. The term r/(s) in Theorem (4-14) is equal to v^s) where rji(s) is the rj-function of the operator d* acting on the space dQ2*"1.. and BOTT. K. F. (3) ATIYAH. P A T O D I AND I. we can formally interpret — n(0) as the ' signature' of the quadratic form Q (on the infinite-dimensional space eZQ24"1). Theorem (4-14) implies of course that. F. Let us define a quadratic form Q on Q2*-1 by (4-21) JY In fact the radical of Q is Ker d. in all odd dimensions. we shall give an alternative proof of these finiteness results independent of Theorems (4-14). Math. This is not always the case but the main results of Thorn's cobordism theory assert that. In other words. 1) 1964. R.

Analytic torsion for complex manifolds. of Math. Sci.. 7 (1971). and SIMONS. F. RAY. Amer. STONG. Advances in Math. 68 (1971). 791—794. and SINGEB. and SINGER. (1971). Sympos. 154-177. J. F. Soc. CONNOB. Pure Math. 288-307. PATODI. Mem. M. H. I. Study 57 (Princeton. J. SEELEY. U. Ann. D. Ann of Math. Nat. 183-281. PATODI. Diff. Curvature and the eigenforms of the Laplace operator. Hilbert modular surfaces. CHEBN. 83 (1966). R. Conduction of Heat in Solids. I. 1965). Proc. 233-249. Spectral asymmetry and Riemannian geometry. Complex powers of an elliptic operator. Geometry 5 (1971). Enseignement Math. second edition (Oxford 1959). S.) HIBZEBBTJCH. Some cohomology classes in principal fibre bundles and their application to Riemannian geometry. Ann. CABSLAW. S. I. K. of Math. 69 M. Notes on Cobordism Theory (Princeton. 145-210. i?-Torsion and the Laplacian on Riemannian manifolds. 229-234. K. S. Soc. 10. R. Hermann (Paris. of Math. F. 129-209.S. London Math. M. Pseudo-differential operators and non-elliptic boundary problems. DB RHAM. V. 1960. Proc. PALAIS. RAY. and SINGEB. Math. J. R. 93 (1971). 19 (1973). I. 20 (1956). The Neumann problem for differential forms on Riemannian manifolds. I (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17) ATIYAH. C. M. and SINGEB. . M. The index of elliptic operators V. V. Seminar on the Atiyah-Singer index theorem. and JAEGEB. Math. P. Ann. Soc. 139-149. D. 98 (1973). Acad. 1968).A. HOBMANDEB. Amer.Spectral asymmetry and Riemannian geometry. 5 (1973). L. G. E. M. Bull. ATIYAH. Variitis Diffirentiables.